Author(s): Gerald Corzo; Michael Siek; Roland Price; Dimitri Solomatine

Linked Author(s): Dimitri Solomatine, Gerald Augusto Corzo Perez

Keywords: Modular models; ANN; Model trees; Hydrological forecasting; Baseflow separation

Abstract: Hydraulic phenomena are composed of a number of interacting sub-processes, so one single model handling all processes is often inaccurate. Modular models allow for modelling subprocesses separately. If data-driven modular models are built, they allow for incorporation of domain knowledge and thus help break down the barriers associated with the "black-box nature" of such models. In this paper we compare two types of modular models that incorporate hydrological knowledge into the modularization process: based on artificial neural networks (ANN), and M5 model trees. The latter have accuracy similar to that of modular models based ANN models, however they can be easier interpreted and are faster. The best performance is obtained from the modular models taking into account the hydrological knowledge.

DOI:

Year: 2007